Attributes

data-fx-chain — the effects

Write these attributes double-quoted, with the JSON’s own quotes as ". The browser reads them through getAttribute and does not care, but scripts/carve.mjs finds them with a name="..." regex, so a single-quoted attribute is invisible to it — the carve reports no existing chain and quietly overwrites work it could not see. & becomes &; nothing else needs escaping.

• Order is signal order. Each node processes what the one before produced.
• type is an effect id from the registry. params are in the units a person thinks in — dB, ms, Hz — and out-of-range values are clamped on read, so a chain that parses is always safe to realise.
• id is a stable handle. Automation addresses nodes by id, never by position, so reordering the chain cannot re-point a lane at a different effect. A node with no id loads fine but cannot be automated. Writing a chain by hand, any unique string works; Studio hands out the first free n1, n2, … so matching that convention keeps a hand-written chain and an edited one looking alike.
• label is what the rack calls this node, replacing the effect’s own name. Write one whenever the node is doing a named job — a chain with two peaking nodes otherwise shows the same row twice and the author cannot tell which is the mud cut and which is the clarity lift. Presets and jobs always set it; a hand-written node should too. See presets.md for the names they use.
• enabled: false is bypass — the node stays in the chain, out of the signal path. Absent means enabled.
• fromCarve: true marks a node the carve analysis generated. Re-running the carve replaces exactly these and leaves hand-built effects alone. Do not set it by hand: a node tagged this way will be deleted by the next carve.

data-automation — the envelopes

• target is volume for the track’s own level, or fx.<nodeId>.<param>.
• t is seconds from the start of the clip, not of the composition. A bed starting at data-start="8" has t: 0 at composition time 8.
• v is in the parameter’s own unit: dB for a gain, Hz for a frequency, 0..1 for volume.
• A lane holds its first value backwards to the start of its clip and its last value forward to the end. So a bed that begins before the voice needs an explicit “no cut” point at t: 0, or it starts out already ducked.
• curve (-1..1) bends the segment leaving a point: positive holds low then rises late. viaX/viaY name an interior point the segment passes through (progress 0..1, value travelled 0..1) and supersede curve when both are present — that is what the timeline writes when a bend is dragged.
• 512 points per lane, maximum.
• A lane whose node is gone is pruned on read rather than erroring.

A lane on a non-automatable parameter is silently inert. Automation is delivered as native AudioParam scheduling, so a knob that no AudioParam backs cannot move: worklet processor options, a WaveShaper curve and a convolution impulse are all set wholesale. fx-registry.md marks each parameter; the four worklet effects (compressor, limiter, gate, bitcrush) have none at all.

data-fx-carve — the carve’s settings

• sources are the element ids of every voice this bed makes room for. They live on the bed being processed, not on the voices. Summed onto the bed’s clock before the analysis, so one set of filters and envelopes covers all of them.
• strength 0..1 derives the whole mechanism (see carveProfile).
• There is no dynamic: a carve always follows the speech.
• enabled is whether the carve applies. It exists because a bed with exactly one candidate voice is carved by default: with “off” represented by an absent attribute, switching it off would read as never-configured and the default would put it back. enabled: false keeps the settings and stops the carve.

This attribute is not read at playback — the chain and lanes it produced are what play. It exists so the settings can be read back and re-derived rather than guessed from the filters, which is what makes changing strength on an existing carve possible.

Older projects may carry the six mechanism numbers (maxCutDb, bands, q, intelligibilityBias, duckDb, headroomDb) instead of strength. They still load: the depth maps back onto a strength and everything else is re-derived. A stored carve with no enabled reads as on, and a single source reads as a one-voice sources list. A stored dynamic is ignored — every carve follows the speech now.